3.1 Design and other characteristics of studies

Included research largely fell into one of four categories: (a) large prospective cohort studies, (b) cohort-nested case-control studies and case-cohort studies, (c) population-based case-control studies and (d) hospital-based case-control studies (Table 1).

There were nine large prospective studies.^{65, 80-85, 125, 126} Three were occupational cohort studies: the Agricultural Health Study (AHS) in the United States,¹²⁵ the AGRICAN cohort study in France,¹²⁶ and the AGRICOH Consortium; the latter is a pooled study combining data from ARGRICAN, AHS and other occupational cohorts, including Cancer in the Norwegian Agricultural Population (CNAP), Australian Pesticide Exposed Workers, Victorian Grain Farmers, Korean Multi-Center Cancer cohort (KMCC), United States Marshfield Epidemiologic Study Area Farm cohort (MESA) and Danish Sund Stald Stud (SUS).^{64, 65} In these cohorts, exposure was assessed at baseline via self-report or through census data; none included biomarkers. The AHS cohort is based in North Carolina and Iowa, and it includes some 57 000 licensed private and commercial pesticide applicators enrolled from 1993 to 1997. Researchers analyzed relationships between binary exposure variables (occupationally exposed vs not exposed), and cumulative and intensity-weighted lifetime days of exposure, a metric that adjusts for factors that influence exposure intensity, for many different pesticides and cancer types.¹²⁵ The AGRICAN cohort is a prospective cohort study in France that enrolled 180 060 current and retired farmers from 2005 to 2007. Pesticide exposure was assessed at enrollment with a questionnaire, and incident cancer was assessed via linkage to population-based registries.¹²⁶

Three studies were general population cohorts, with occupation being used to assess exposure through a job exposure matrix (JEM); a JEM is a tool that classifies individuals into likely exposure categories based on time spent in various occupational roles.¹²⁷ The studies included the International Childhood Cancer Cohort Consortium⁸³ (pooled birth cohorts), a Swiss census-based cohort study,⁸⁴ and the Canadian Census Health and Environment Cohort.⁸⁵ In these studies, subjects occupationally exposed to pesticides were compared to those nonoccupationally exposed.^{83, 84}

Two of the prospective cohort studies measured dietary exposure to pesticides using food frequency questionnaires for fruit and vegetable intake, and information on organic status, and probable pesticide residue levels from government agricultural databases on farming practices.^{80, 81} Only one prospective cohort study used biomarkers for exposure assessment.⁸²

In contrast to the cohort studies, nested case-control and case-cohort studies typically included prospectively measured biomarkers for exposure assessment (eg, biomarkers measured in biological samples collected at study entry, years before cancers occurred). Some such studies used questionnaires as well; questionnaires and biomarkers often complement each other. These studies are a part of larger cohort studies and, per design, were highly efficient (eg, precise) with small to medium sample sizes (median sample size ~300 participants). Examples include case-control studies nested within the Janus Serum Bank Cohort,¹²⁸ a Norwegian population-based biobank established in 1973, and the Danish National Birth Cohort, a population-based birth cohort out of Denmark established in 1996¹²⁹ (Table 3).

Population-based case-control studies^94-112 had medium to large sample sizes (median sample size ~1800 individuals). They used population-based data sets from census, and cancer and birth registries. In these studies, exposure was typically assessed via existing data sets, using occupation and job-exposure matrices, or via geographical analysis of proximity to agricultural sites as a proxy for exposure; at other times exposure was self-reported using self-administered or interviewer-administered questionnaires (Table 4).

All hospital-based case-control studies used biomarkers for exposure assessment. Pesticides were quantified in blood samples,^{113, 114, 116-119} or in adipose tissue samples (in breast cancer)¹¹⁵ if surgery was performed. Less than 8% of the studies assessed whether varying levels of pesticide exposure influenced differences in survival among patients with cancer (Tables 5 and 6).^121-124

About 10% of the studies assessed prenatal and childhood exposures and the development of cancer in children, while the adult studies typically studied populations in their 50s and 60s. Thirty percent of the studies included only women, 5% of the studies included only men, and 65% of the studies lumped together both genders; several of the latter were occupational studies, which had a large proportion of males.¹²⁵ Forty-six percent of the studies assessed occupational exposures. A slightly larger proportion of reports looked at agricultural exposures (including residences near farms). Only six (9.5%) of the reviewed studies specifically assessed residential pesticide use.^{78, 100, 105, 107, 109, 121} About 16% of the studies used prospectively measured biomarkers^{19, 82, 86-88, 90-93, 124} (Table 1).

3.2 Main empirical findings

The main empirical findings from the included studies are summarized below and in Tables 2-6 grouped by study design. Two studies from the AGRICOH consortium were published during the time period of interest. Togawa and colleagues examined the incidence of primary cancer diagnosis and specific subtypes in the occupational cohorts compared to the general population; in women, they found increased risks for melanoma of the skin (meta-SIR = 1.18, CI: 1.01-1.38) and multiple myeloma (meta-SIR = 1.27, CI: 1.04-1.54), and in men, for prostate cancer (meta-SIR = 1.06, CI: 1.01-1.12). They also observed a deficit for the incidence of cancers of the bladder, breast, colorectum, esophagus, larynx, lung, pancreas, and all cancers combined (meta-SIR for all cancers combined = 0.83, 95% CI: 0.77-0.90).⁶⁴ The other AGRICOH study specifically investigated non-Hodgkin's lymphoma (NHL), and observed an increased risk for NHL among ever users of terbufos (meta-HR = 1.18, CI: 1.00-1.39); for chronic lymphocytic leukemia/small lymphocytic lymphoma among ever users of deltamethrin (meta-HR = 1.48, CI: 1.06-2.07); and for diffuse large B-cell lymphoma in ever users of glyphosate (meta-HR =1.36, CI: 1.00-1.85; Table 2).⁶⁵

In an analysis from the AHS focused on thyroid cancer, an increased risk of this cancer was found with exposure to metalaxyl (hazard ratio (HR) = 2.03, confidence interval (CI): 1.16-3.52), and lindane (HR = 1.74, CI: 1.06-2.84). High usage of the insecticide carbaryl was inversely associated with thyroid cancer (HR = 0.20, CI: 0.08-00.53).⁷² Another AHS analysis that examined renal cell carcinoma (RCC) found, in an unlagged analysis, an increased risk of RCC among ever users of 2,4,5-T compared to never users (RR = 2.92, CI: 1.65-5.17).⁷³ In a 20-year lagged analysis, exposure-response relationships were observed between RCC risk and chlorpyrifos (RR = 1.68, CI: 1.05-2.70), chlordane (RR = 2.06, CI: 1.10-3.87), atrazine (RR = 1.43, CI: 1.00-2.03), cyanazine (RR = 1.61, CI: 1.03-2.50), and paraquat (RR = 1.95, CI: 1.03-3.70).⁷³ No associations were observed with glyphosate exposure and solid tumors or lymphoid malignancies overall, nor with non-Hodgkin's lymphoma. An increased risk of acute myeloid leukemia (AML) was observed among the group with highest exposure to glyphosate (RR = 2.44, CI: 0.94-6.32).⁶¹ Other analyses from the AHS suggest that the risk of aggressive prostate cancer is increased by dimethoate use (HR = 1.37, CI: 1.04-1.8) and decreased with triclopyr use (HR = 0.68, CI: 0.48-0.95; Table 2).⁷⁵

The GraMo cohort from Southern Spain was the only prospective cohort study included in this review that used biomarkers for exposure assessment. It found an increased risk of nonhormone-dependent cancer for individuals with increased exposure to β-hexachlorocyclohexane (β-HCH) (HR = 1.70; CI: 1.09-2.64) and hexachlorobenzene (HCB; HR 1.54, CI: 1.02-2.33).⁸²

The prospective cohort studies that looked at potential dietary exposures to pesticides in relation to cancer report conflicting findings. Sandovali-Insuati and colleagues found no association between high-pesticide food intake and risk of cancer overall or for specific subtypes.⁸⁰ However, Reboulliat et al., in a study looking at pesticide mixtures using Principal Component Analysis (PCA) reported increased risk of breast cancer for overweight and obese women with a higher component 1 score, reflecting high exposures to chlorpyrifos, imazalil, malathion, and thiabendazole (HR_Q5 = 4.13, CI: 1.50-11.44). They also reported an inverse association between breast cancer risk and component 3 score, reflecting low exposure to synthetic pesticides, and [HR_Q5 = 0.57, CI: (0.34-0.93); Table 2].⁸¹

As mentioned, two large, prospective studies looked at the relationship between parental exposure to pesticides, assessed through job-exposure matrices and childhood cancer. In the census-based cohort study in Switzerland, maternal and paternal pesticide exposure increased noncentral nervous system (CNS) solid tumors (high vs none, father: adjusted HR = 1.84, CI: 1.31-2.58; mother: HR = 1.79, CI: 1.13-2.84).⁸⁴ No associations were found between parental exposure and leukemia, lymphoma or CNS tumors (Table 2).⁸⁴

The similarly designed analysis of pooled data on more than 329 000 participants from the International Childhood Cancer Cohort Consortium indicated that paternal exposure to pesticides increased the risk of childhood AML (herbicides, HR = 3.22, CI: 0.97-10.68; insecticides, HR = 2.86, CI: 0.99-8.23), but not of acute lymphoblastic leukemia (ALL) or CNS tumors (Table 2).⁸³

Several nested case-control studies analyzed data from the aforementioned Janus cohort, a population-based biobank with sera samples prospectively collected from 1972 to 2004 from more than 300 000 Norwegian men and women.⁹⁰ In a study assessing the relationship between organochlorine pesticides and thyroid cancer, an inverse relationship between DDT metabolites and thyroid cancer risk was observed (OR = 0.80, CI: 0.66-0.98).⁹² Additionally, significant positive exposure-response relationships were seen between chlordane metabolites and thyroid cancer among participants born from 1943to 1957. These findings are compatible with birth cohort effects and with effects specific to certain historic periods.⁹² Investigating AML, researchers observed significant exposure-response relationships for heptachlor epoxide (third vs first tertile OR = 2.85, CI: 1.05-7.73) and dieldrin (third vs first tertile OR = 2.71, CI: 1.07-6.83) with AML risk.⁹¹ No statistically significant exposure-response associations with AML risk were observed for other pesticides, although the magnitude of the estimates was substantial for chlordane/heptachlor metabolites (third vs first tertile OR = 2.26, CI: 0.91-5.63) and p,p'-DDT (OR = 2.09, CI: 0.83-5.26; Table 3).⁹¹

Another study included analyses from two case-control studies nested in different cohorts. Relationships between serum concentrations of organochlorine pesticides and primary liver cancer were assessed separately among the Janus cohort and a Northern Californian Kaiser Permanente cohort in one report.⁹⁰ Among Janus participants, an exposure-response relationship was observed for p,p′-DDT (second and third tertile ORs = 1.70, CI: 0.66-4.38, and 2.14, CI: 0.79-5.75, respectively). Among the Californian participants, there was an exposure-response relationship for trans-nonachlor (second and third tertile ORs = 1.63, CI: 0.87-3.06 and OR = 1.95, CI: 0.98-3.86), as well as an elevated risk for the highest tertile of oxychlordane (OR = 1.87). None of these relationships were statistically significant (Table 3).⁹⁰

A case-control study nested within the European Prospective Investigation into Cancer and Nutrition (EPIC) cohort assessed the relationship between pesticide exposure and pancreatic cancer risk. Increased risk was observed at higher concentrations of p,p′-DDT, trans-nonachlor (OR upper quartile = 1.55, CI: 1.06-2.26), β-HCB and the sum of six organochlorine pesticides (OR upper quartile = 1.48, CI: 1.00-2.20; Table 3).⁸⁶

In a case-cohort study nested within the Korean National Cancer Center Community (KNCCC) cohort, a significantly increased risk of colorectal cancer (CRC) was associated with serum concentrations of cis-heptachlor epoxide (third tertile HR = 2.76, CI: 1.25-6.07); and significant exposure-response relationships were observed for trans-nonachlor (second tertile HR = 3.90, CI: 1.56-9.75, third tertile HR= 4.86, CI: 1.95-12.16) and p,p′-DDD (second tertile HR = 6.02, CI: 2.05-17.70, third tertile HR = 7.43, CI: 2.42-22.84).⁸⁷ A different case-cohort study out of Korea looking at prostate cancer found evidence of association between any measured pesticides and prostate cancer.⁹³ A pilot case-control study including Hawaiian participants that provided urine samples from the Multiethnic Cohort study showed a nearly 4.5-fold increased risk of developing breast cancer in the highest vs lowest quintile of AMPA (a glyphosate metabolite) excretion: Q5 vs Q1 OR = 4.49, CI: 1.46-13.77 (Table 3).⁸⁸

Four population-based case-control studies investigated the relationship between self-reported use of pesticides at home among mothers, and childhood cancer. The Costa Rican Childhood Leukemia study observed significant associations between maternal insecticide use inside the home in the year before pregnancy, during pregnancy, and while breastfeeding; use during such three periods was associated with increased risk of ALL among boys: adjusted OR = 1.63 (CI: 1.05, 2.53), 1.75 (1.13, 2.73) and 1.75 (1.12, 2.73), respectively.¹⁰⁵ They also observed dose-response relationships between the frequency of pesticide use in the home in the year before pregnancy, during pregnancy, and while breastfeeding and ALL among boys and girls: OR comparing high to low = 1.56 (CI 1.07-2.27), 1.58 (1.08-2.31) and 1.56 (1.07-2.29), respectively (Table 4).¹⁰⁵ In the ESTELLE study, a French nationwide population-based case-control study, maternal use of pesticides during pregnancy was associated with the risk of Wilms' tumor in children (OR = 1.6, CI: 1.1-2.3).¹⁰⁰ Among insecticides, specifically, the association was similar: OR = 1.7, CI: 1.1-2.6; when insecticides were used more than once a month, the OR was 1.9 (CI: 1.2-3.0).¹⁰⁰ The two other studies report pooled data from the ESTELLE study and the ESCALE study, another French nationwide population-based case-control study. One showed an association between maternal residential use of pesticides during pregnancy and childhood brain tumors (OR = 1.4, CI: 1.2-1.8). The other reported more specifically on neuroblastoma (NB), and found an increased risk of NB among children prenatally exposed to any type of pesticide (OR = 1.5, CI: 1.2-1.9; Table 4).^{107, 109}

All eight hospital-based case-control studies used biomarkers (Table 5). Sample sizes were small, especially in two instances.^{116, 117} All such studies reported more than one positive association. Notably, both Lee et al¹¹⁸ and Abolhassani et al¹¹⁷ found relationships between several serum concentrations of pesticides and colorectal cancer. Also Huang,¹¹⁵ Wielsoe,¹¹⁹ Mekonen¹¹³ and He,¹²⁰ all report statistically significant associations between dichlorodiphenyldichloroethylene (DDE) concentrations and breast cancer.

Three survival studies assessed the survival of patients after a diagnosis of breast cancer, while one assessed survival after a diagnosis of NHL (Table 6). The only of the breast cancer studies to use a biomarker for exposure assessment, and to control for confounding variables¹²² found an increased risk of all-cause mortality among the highest vs lowest DDE tertile and the highest tertile vs nondetectable DDT (HR = 1.95 and 1.64, respectively). It also reported an increased risk of 20-year conditional breast cancer-specific mortality among women overall (HR = 1.69), higher for black women (2.36), than for white women (1.57; Table 6). Concentrations in adipose tissue of polychlorinated biphenyls (PCBs) and organochlorine pesticides were not associated with NHL survival.¹²¹

4.1 Temporality

Thirty-three of the 63 studies identified for this review assessed exposure to pesticides prospectively, prior to cancer development.^{19, 61, 64-93} These studies provide the strongest evidence due not only to their prospective and precise exposure assessment, but also due to large sample sizes. Additionally, conversely to case-control and cross-sectional studies, these studies are less susceptible to recall bias, a differential misclassification that tends to bias these studies away from the null.⁶³ Among these 33 studies, including studies reporting consistent associations of increased risk of AML^{61, 87, 91} and CRC, the causal condition of temporality is fulfilled. Most of these studies report positive associations between pesticide use and cancer. The largest of these studies are from the AGRICOH consortium; pooled analyses of several prospective cohorts. The studies describe increased risk of NHL associated with terbufos, an organophosphate pesticide with similar chemical properties to glyphosate, and increased incidence of MM and CNST compared to the general population.^{64, 65} Additionally, studies that used biomarkers prospectively were also likely to report positive associations. The EPIC cohort, with robust covariate adjustment and sensitivity analyses, reported increased risk of pancreatic cancer with exposure to summed organochlorine pesticides, β-HCH, DDT, and trans-nonachlor.⁸⁶

4.2 Consistency

The existing literature is not well suited to address this consistency criterium, as the studies are highly variable in exposure assessment, dose, timing, latency periods and so forth. Thus, here we describe patterns observed that provide some evidence of consistency and discuss potential reasons for heterogenous findings. Though all studies tested multiple hypotheses and thus some are likely due to chance, all but 6 of 63 studies reported at least one statistically significant positive association between some measure of exposure to pesticides and cancer. Several studies report positive relationships that are robust, with consistent results in sensitivity analyses.^{19, 61, 66, 67, 72-74, 83, 84, 86, 89-92} Further, these studies often assess exposure to several, highly correlated, pesticides. Some analyses were conducted for individual chemicals, sometimes with adjustment for others.^{66, 125} Sometimes this may be a form of overadjustment, with little significance both from mechanistic and practical standpoints. Adjusting for exposure to correlated pesticides may also introduce multicollinearity, and result in potentially biased and unstable estimates of association.¹³⁴

While study conclusions varied, four out of six studies assessing AML reported evidence of increased risk for subjects with higher exposure to specific pesticides; particularly glyphosate, hepatachlor-epoxide, dieldrin and chlorpyrifos and carbamates.^{61, 76, 83, 91, 101, 112} In Andreotti et al, this risk is insignificant for unlagged models, but significant when considering a 20-year lagged analysis.⁶¹ The studies reporting positive associations include large, prospective studies that measured several individual pesticides; one included prospectively collected biomarkers for exposure ascertainment. In the two exceptions, pesticides overall (rather than specific pesticides) were analyzed; this approach is documented to have weaker associations with AML.¹¹² Additionally, Coste et al discuss limitations in their analysis, as only three AML cases were in the high exposure category.⁸⁴ Results were also consistent across studies looking at CRC: all three studies observed an increased risk of CRC with higher exposure to pesticides.^{87, 117, 118} It is also worthwhile to note that (a) these studies provide evidence of causal relationships between pesticides and different types of cancer, (b) these studies look at pesticides with different mechanisms of action, and (c) to fully conclude that the consistency criteria was fulfilled, more research would be needed on specific individual pesticide-cancer relationships.

Although the body of evidence has increased recently, there are still issues of reproducibility and conflicting findings. One well-known reason for these inconsistencies is the existence of a long latency period for cancer development. IARC states that “experience with human cancer indicates that the period from first exposure to the development of clinical cancer is sometimes longer than 20 years; latent periods substantially shorter than 30 years cannot provide evidence for lack of carcinogenicity”.¹¹ Even cohort studies with the longest follow-up periods, like the AHS, have a roughly 20-year follow-up period; they may thus fail to capture longer-term effects. They may also fail to assess the effects of exposure during critical windows.¹²⁵

Another possible explanation of the apparently conflicting evidence is heterogeneity in unmeasured co-exposures and in exposure (mis)classification.¹³⁵ While some studies measured specific pesticides, many measured exposures generally or used various classification schemes to operationalize exposure variables for statistical models. Examples of exposure variables include ever/never use of specific pesticides, cumulative intensity-weighted lifetime days, occupationally exposed vs not, proximity to facilities using pesticides, and serum/tissue levels. While all such practices are acceptable, they complicate comparisons of findings across studies; and such difficulties should not be equated with lack of consistency.¹³⁵ However, to fully address this criterium, more research replication efforts are needed.

4.3 Strength

As cancer is multifactorial, it is not expected for the association between cancer and any individual environmental exposure to be large in magnitude. However, we can assess the strength of an association in response to increasing dosage to assist in inferring causality. Several of the studies identified exposure-response relationships between pesticide exposure and cancer risk in humans of a substantial strength or magnitude.^{73, 86, 87, 90-92} These relationships show that increased exposure granularly increases risk, strengthening the claim that the relationship is in fact causal.

4.4 Plausibility

Several mechanisms have been proposed to explain the observed relationships between cancer and pesticides. Among these are genotoxicity, hormone disruption (particularly acetylcholine esterase inhibition),¹³⁶ oxidative stress, inflammation, immune modulation and procarcinogen activation.^11-14 Oxidative stress, a main proposed mechanism, has been connected to cytochrome P-450 induction, resulting in the formation of reactive oxygen species with the ability to cause DNA damage, genome instability and cell proliferation.¹³⁷

4.5 Coherence

The relationship between exposure to many pesticides and the occurrence of cancer in animal and in vitro studies is well established; it is often summarized in the IARC monographs, for instance.^{11, 40, 136, 138-140} In the primary human studies we reviewed here, there is modest coherence, partly due to the difficulties with exposure imprecision. A major conclusion of this article is that newer, biomonitoring-based designs can resolve some of the issues with coherence in the literature.

4.6 Specificity

Studies have identified some stronger effects in aging populations, but this is not what is generally considered a specific population. Indeed, in cancer research, specificity is an uncommon causal property: few exposures have unique effects and few cancers are due to exclusive exposures. Accordingly, few studies on pesticides and cancer have assessed specificity directly, such as by identifying control conditions not associated with pesticide exposure. Nevertheless, future studies could sometimes consider specificity; for example, ask whether their findings suggest a specific population with the effects or a specific site of the disease with no other likely explanation. Studies could also assess “falsifier exposures” (exposures tested to confirm specificity of effect of pesticides vs an exposure without biological plausible carcinogenicity), and use “negative controls”⁴⁰; for example, examine associations of pesticides with conditions known to be unrelated to the pesticides. There may also be specific routes of pesticide exposure (eg, dermal, inhalational) that may be most problematic.

4.7 Analogy

The existence of causal relationships between similar chemicals and cancer increases the likelihood that observed associations are causal. There are several instances of similar environmental chemicals being carcinogenic. For example, formaldehyde, a substance commonly used as a germicide, is known to cause myeloid leukemia, and rare cancers including cancers of the nasal cavity, and nasopharynx.¹⁴¹ Also, and most notably, ethylene oxide, a chemical used as a pesticide, is recognized as a cause of lymphoma and leukemia.¹⁴²

4.8 Further suggestions for future work

To improve the validity and relevance of the available knowledge regarding exposure to pesticides and cancer risk in humans, studies must increase in number, diversify and improve in several respects. We need more large population cohort studies that allow for longer latency periods. These cohorts also enable to conduct efficient nested case-control studies using biological samples (collected at baseline) for exposure assessment. More hypotheses need to be tested in younger and nonoccupational cohorts, with emphasis on nonoccupational sources of exposure (as diet) and on developmental periods of susceptibility.

The strengths of the main study designs (large prospective cohort studies, cohort-nested case-control studies with long follow-up) hence deserve greater allocation of funds and wider recognition in regulatory institutions, companies, clinical organizations and the media.

Tools from molecular epidemiology (including environmental, epigenetic, metabolomic, exposomic and proteomic measurements used in population-based studies) can be leveraged in cancer cohorts,¹⁶ as shown, for instance, in recent papers from the European Prospective Investigation into Cancer and Nutrition (EPIC) study,^{86, 143} or in the Prostate, Lung, Colorectal and Ovarian Cancer Screening Trial (PLCO).¹⁴⁴ Also recent is the integration of multi-omics analysis in cohorts as the Norwegian Women and Cancer (NOWAC),¹⁴⁵ and the EPICURE study.¹⁴⁶ Additionally, advances in biomonitoring will also contribute to progress.¹⁴⁷ Such technologies include wearable personal sensors, location trackers and nontargeted methods.^{148, 149} As cohorts incorporating such technologies mature, evidence will increase in quality and strength as exposure assessment will be a more accurate reflection of real-world exposure. Additionally, multi-omics data will provide rich biological context for mechanisms of environmental carcinogens. Successful adaptation of these techniques and tools remains both a challenge and an avenue for future research.^150-152

Furthermore, humans are rarely subjected to individual environmental exposures, but rather to complex chemical mixtures, and novel methods have been developed to allow for analysis of these mixtures.¹⁵³ Low doses of even noncarcinogenic chemicals can act synergistically to contribute to carcinogenesis, and mixtures analysis has long been a concern for researchers.^{154, 155} In this review there were only three studies that examined interactions of pesticides and other factors in cancer, and no studies that employed statistical methods designed to assess effects of environmental chemical mixtures. More research on chemical-chemical, chemical-(epi)genetic and other interactions and mixtures¹⁵⁶ must be pursued.

Existing evidence is largely from older, occupationally or agriculturally exposed populations; sometimes, studies also used exposure assessment methods with important limitations. While some large prospective cohort studies were conducted, these studies were entirely occupational, not population-based, and relied on self-report of specific pesticide use or crop management to assess exposure, often years after exposure took place.^{66, 72} These studies are more susceptible to measurement error than studies using biomarkers, and the occupational cohorts cannot be generalized to much of the population at large. Studies that used biomarkers to assess exposure were mostly small hospital-based case-control studies, which collect biological specimens (where exposures are later determined) at the time of diagnosis, often months after the disease is progressing subclinically; hence, they are susceptible to disease progression bias, a form of reverse causality (concentrations of biomarkers may be a result of the disease, not a cause).^{40, 86} Such studies may also suffer from bias in control selection, particularly when studying gene-environment interactions.^{135, 157, 158}

Of the existing study designs, the design that often balances best exposure assessment methods with sample size and bias control is the cohort-nested case-control study. Unfortunately, only a small number of studies were published in the time period of interest and represent only 15% of the included studies. These studies benefit from prospectively collected biological samples that cannot be influenced by disease status and integrate all sources of exposure (including food, water or air if appropriate). They are also a part of a larger cohort, which contributes to cost and logistical efficiency.¹³⁵ Cases and controls are matched on follow-up time and some potential confounders to efficiently control for confounding and biomarker storage/batch effects (eg, cases and controls are matched on fasting status at blood collection).^{86, 159} Such study design, while currently underrepresented in the literature, is growing in popularity and is a good option to strengthen the current body of evidence on the relationship between cancer and pesticides.¹³⁵

The current body of research is also lacking representation of urban cohorts and data on residential and food exposure to pesticides. Only 10.5% of the reviewed studies looked at residential pesticide use, and there were no explicit investigations of urban pesticide exposure. While frequently overlooked, cities are an important landscape of pesticide use. The NYC Department of Health and Mental Hygiene released a report in 2017 documenting a total of 6711 gal of liquid and 163 182 pounds of solid pesticide products across city agencies the previous year. This large quantity only represents only a minority of urban pesticide use; comparisons to state commercial pesticide use data revealed that city agency use only accounts for an estimated 3% of liquid pesticides, and 21% of the solid pesticides applied in the city.¹⁶⁰ The level of pesticide use in cities is thus nontrivial, and different soil properties, environmental fate of pesticides and exposure profiles in urban environments warrant further investigation of the unique health burden of urban pesticide use.¹⁶¹

4.9 When is enough evidence enough?

While further research is always needed, the question ultimately is whether the evidence is sufficient to better regulate pesticides on the basis of carcinogenicity. Debates about causality and the sufficiency of evidence to act not new; for instance, causal inference was addressed by Immanuel Kant and other philosophers centuries ago. It was also raised decades ago by Hulka and others in the context of carcinogenesis.¹⁶² It applies both to sufficient evidence to act, when a causal relationship is likely, and to not act, when a causal relationship is unlikely. Hill also emphasized that “all scientific work is incomplete… [and] liable to be upset or modified by advancing knowledge.”¹³³ He added: “That does not confer upon us a freedom to ignore the knowledge we already have or postpone the action that it appears to demand at a given time.”¹³³ Although there are limitations and gaps in the research as well as some discordant findings, there is sufficient evidence for action to regulate pesticides based on their carcinogenicity.

Science has also evolved greatly over the near 60 years since Hill gave his landmark lecture, and scientific studies and institutions have addressed crucial flaws in the risk assessment paradigm leveraged in the European Union, the United States and elsewhere to manage environmental exposures. The conventional paradigm sometimes assumes, in particular, a linear dose-response, which has been shown not to occur in many causal relationships.^{40, 163} Indeed, the endocrine system, in particular, has revealed U-shaped (nonmonotonic) and nonlinear exposure-response functions, including for pesticides and developmental neurotoxicity.^164-166

Conflicting findings and gaps in data have tangible consequences. In 2016, the United States Environmental Protection Agency (US EPA) published an issue paper stating that glyphosate (which, as mentioned earlier, is classified by IARC as probably carcinogenic), is not a carcinogen; and it decided to allow for its continued and increasing use, a decision EPA reaffirmed in 2020.^{167, 168} The European Food Safety Authority came to a similar conclusion as the EPA in a 2015 report, where they claim glyphosate is “unlikely to pose a carcinogenic threat to humans”.¹⁶⁹ The main reason for the opposed conclusions from these authorities is that IARC studied the genotoxic and carcinogenic properties of not only pure glyphosate, but also of glyphosate-based formulations, pesticides where glyphosate is the active ingredient.¹⁷⁰ The evidence of carcinogenicity for glyphosate-based formulations and metabolites is stronger than for technical glyphosate, as co-formulants affect absorption, distribution, metabolism and excretion of glyphosate.^{170, 171} Additionally, the IARC committee only considers peer-reviewed literature, while the assessments done by EPA and EFSA also considered several industry studies conducted by private companies that register the chemicals in question. In response to the ongoing debate, calls have been made within the scientific community for more independently conducted epidemiological research, particularly among the highest exposed and the most vulnerable populations, as well as for more comprehensive and transparent evaluations by independent agencies.^{170, 172, 173}